The ensemble learning approach based on random sampling or cluster sampling was developed to predict the number of desulfurization tower circulating pumps opened. The database was constructed from a realistic power plants, the outliers and unsteady values in the initial one were deleted, and the input parameters of the ensemble learning model with high correlation coefficients with the output were extracted. The problems of imbalanced samples in classification, missing evaluation criteria for optimal samples and desulfurization optimization were solved. Results showed that the improved ensemble learning model had a 33% increase in overall prediction accuracy compared with the original model. In addition, the cluster sampling was slightly better than the random sampling. Furthermore, the recall of a single category prediction was analyzed, and the recall values of different algorithms for the minority category and the majority category were compared. Results showed that the two improved sampling methods had greatly improved the minority category prediction, and the recall reached more than 90%, besides it also had certain improvement on the majority. Finally, the difference in sample distribution and model accuracy was discussed when the pump combination was used as the model’s output.